Optical imaging systems often suffer from wavefront distortions introduced by aberrations, which degrade image quality and limit performance in fields such as astronomy, microscopy, and long-range imaging. Traditional wavefront sensing methods require dedicated hardware or controlled illumination conditions, posing challenges for deployment in dynamic or resource-constrained environments. This thesis presents two computational imaging frameworks that enable high-quality imaging through unknown aberrations using learning-based techniques, without requiring guidestars, paired training data, or specialized sensors.

The first contribution is a guidestar-free wavefront correction framework that leverages asymmetric apertures and two neural networks to estimate and correct phase aberrations from extended natural scenes. By exploiting the injective mapping properties of asymmetric apertures, such as triangles, the proposed method breaks conjugate flip ambiguities inherent in conventional phase retrieval. Experimental results show that this approach achieves over 9 dB improvement in PSNR compared to symmetric apertures and effectively corrects unknown isoplanatic aberrations.

The second contribution is NeRT, an unsupervised turbulence mitigation framework based on implicit neural representations. NeRT decomposes turbulence effects into spatially and temporally varying tilt and blur components, following a tilt-then-blur model. It jointly learns grid deformation, a coordinate-based image generator, and a shift-varying blur module to recover a clean image from distorted observations without ground truth supervision. NeRT demonstrates strong generalization across atmospheric and water turbulence datasets and outperforms existing supervised and unsupervised methods on both synthetic and real-world scenes.

Together, these contributions offer practical solutions for imaging through complex aberrations, pushing the boundarieses of adaptive optics and computational imaging in uncontrolled environments.